Run agent-browser on your computer

Read websites, interact with pages and capture screenshots in a separate local browser.

Give your agent a working browser

Your agent gets agent-browser’s MCP tools for opening pages, reading their contents and interacting with buttons and forms. The browser stays open between tools for the life of that connection.

For a first result, supply a page address and get its screenshot and readable accessibility snapshot. This browser has its own session: it does not inherit your usual browser’s accounts, extensions or saved passwords.

You bring

a JSON object with url and optional width, height, full_page and wait_selectorpage.json

You get

a screenshot at the selected viewport or full page heightpage.png
the page's readable accessibility treesnapshot.txt
the page address, title, capture options and output checksumspage.json
agent-browser's navigation, page interaction and capture tools; browser state lasts for this MCP connection

Licences & source

agent-browser Apache-2.0, with bundled notices; Chromium and its Debian dependencies retain their distribution notices. UGT's driver and checks are Apache-2.0.

Setup & verification

agent-browser at commit e7519957b906

Run on clean x86-64 (Intel/AMD) and ARM64 machines on 10 Nov 2026. Its check passed on a known sample and rejected deliberately broken results.

This checks the job this page describes, not every feature of the original project.

Known issues

This is a headless CPU browser, not a connection to your signed-in desktop browser. Sites may require login or block automation.
Closing the MCP connection ends this browser process. Do not use it as persistent storage for an account or assume a new connection resumes an earlier page.
Chromium's sandbox is disabled inside the container. UGT's container boundary still applies; do not expose this MCP server as an unauthenticated public service.

Recipe details